Instructions to use summerMC/Looped-Qwen-pre with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use summerMC/Looped-Qwen-pre with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="summerMC/Looped-Qwen-pre", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("summerMC/Looped-Qwen-pre", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use summerMC/Looped-Qwen-pre with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "summerMC/Looped-Qwen-pre"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "summerMC/Looped-Qwen-pre",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/summerMC/Looped-Qwen-pre

SGLang

How to use summerMC/Looped-Qwen-pre with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "summerMC/Looped-Qwen-pre" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "summerMC/Looped-Qwen-pre",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "summerMC/Looped-Qwen-pre" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "summerMC/Looped-Qwen-pre",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use summerMC/Looped-Qwen-pre with Docker Model Runner:
```
docker model run hf.co/summerMC/Looped-Qwen-pre
```

Looped-Qwen-pre / README.md

summerMC

Looped GDN: weights + custom code + scripts

51ad11c verified 15 days ago

preview code

Raw History Blame Contribute Delete

2.84 kB

---
license: apache-2.0
library_name: transformers
base_model:
- j-llm/Qwen3.5-2B-SpeedX2
- Qwen/Qwen3.5-2B
tags: [gated-deltanet, linear-attention, looped-transformer, weight-sharing, distillation, custom_code]
language: [ja, en]
pipeline_tag: text-generation
---
# summerMC/Looped-Qwen-pre

[`j-llm/Qwen3.5-2B-SpeedX2`](https://huggingface.co/j-llm/Qwen3.5-2B-SpeedX2) (全層 Gated DeltaNet) の 24 層を
**空間マージ**して重み共有し、**2 重ループ**にした Looped GDN (反復ステップ毎 LoRA 付き)。教師 [`Qwen/Qwen3.5-2B`](https://huggingface.co/Qwen/Qwen3.5-2B) から蒸留。

```
for t in range(6):            # 外側ループ
    h = RoPE注入(h)             # 位置の再注入 + ループ段埋め込み
    for k in range(loops[t]):   # 内側ループ (ゲート付き)  loops = [4, 4, 4, 4, 4, 4]
        h = h + g_k(h) * (GDN_t(h) - h)   # GDN_t = 元の 4 層を平均マージした共有ブロック
```

## 使い方 (transformers)

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "summerMC/Looped-Qwen-pre"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True, dtype=torch.bfloat16).cuda().eval()
ids = tok.apply_chat_template([{"role": "user", "content": "日本の首都はどこ？"}], add_generation_prompt=True,
                              return_tensors="pt", return_dict=True, enable_thinking=False)["input_ids"].cuda()
out = model.generate(ids, max_new_tokens=128, do_sample=True, temperature=0.7, top_p=0.9, repetition_penalty=1.1)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))
```

`pip install -U transformers flash-linear-attention` 推奨 (GDN 高速カーネル)。

## ファイル
| ファイル | 内容 |
|---|---|
| `modeling_looped_gdn.py` / `configuration_looped_gdn.py` | transformers 用カスタムコード (`trust_remote_code`) |
| `looped_gdn.py` | モデル本体・空間マージ変換 (`python looped_gdn.py --src j-llm/Qwen3.5-2B-SpeedX2 --groups 6`) |
| `scripts/colab_looped_distill.py` | 蒸留スクリプト (Colab に貼るだけ) |
| `scripts/infer_looped_gdn.py` | 推論 / 対話スクリプト |
| `scripts/colab_ast_transplant.py` | 高度な空間移植 (活性考慮 RegMean + DP 可変長分割 + 活性考慮 LoRA) |
| `scripts/test_looped_gdn.py` | 単体テスト |
| `looped_config.json` | 旧形式の設定 (`LoopedGDNForCausalLM.load` 用) |

## 学習
- Stage 1: グループ単位 hidden 蒸留 (相対 MSE + コサイン) / Stage 2: 教師 logits KL + hidden 損失
- データ: C4 (日本語・英語)

## 注意
蒸留による近似で教師より性能は低下します。ライセンスは元モデル (Qwen/Qwen3.5-2B) に従います。
元モデル: Qwen Team / 全 GDN 化: summerMC, j-llm

	---
	license: apache-2.0
	library_name: transformers
	base_model:
	- j-llm/Qwen3.5-2B-SpeedX2
	- Qwen/Qwen3.5-2B
	tags: [gated-deltanet, linear-attention, looped-transformer, weight-sharing, distillation, custom_code]
	language: [ja, en]
	pipeline_tag: text-generation
	---

	# summerMC/Looped-Qwen-pre

	[`j-llm/Qwen3.5-2B-SpeedX2`](https://huggingface.co/j-llm/Qwen3.5-2B-SpeedX2) (全層 Gated DeltaNet) の 24 層を
	空間マージして重み共有し、2 重ループにした Looped GDN (反復ステップ毎 LoRA 付き)。教師 [`Qwen/Qwen3.5-2B`](https://huggingface.co/Qwen/Qwen3.5-2B) から蒸留。

	```
	for t in range(6): # 外側ループ
	h = RoPE注入(h) # 位置の再注入 + ループ段埋め込み
	for k in range(loops[t]): # 内側ループ (ゲート付き) loops = [4, 4, 4, 4, 4, 4]
	h = h + g_k(h) * (GDN_t(h) - h) # GDN_t = 元の 4 層を平均マージした共有ブロック
	```

	## 使い方 (transformers)

	```python
	import torch
	from transformers import AutoModelForCausalLM, AutoTokenizer
	repo = "summerMC/Looped-Qwen-pre"
	tok = AutoTokenizer.from_pretrained(repo)
	model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True, dtype=torch.bfloat16).cuda().eval()
	ids = tok.apply_chat_template([{"role": "user", "content": "日本の首都はどこ？"}], add_generation_prompt=True,
	return_tensors="pt", return_dict=True, enable_thinking=False)["input_ids"].cuda()
	out = model.generate(ids, max_new_tokens=128, do_sample=True, temperature=0.7, top_p=0.9, repetition_penalty=1.1)
	print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))
	```

	`pip install -U transformers flash-linear-attention` 推奨 (GDN 高速カーネル)。

	## ファイル
	\| ファイル \| 内容 \|
	\|---\|---\|
	\| `modeling_looped_gdn.py` / `configuration_looped_gdn.py` \| transformers 用カスタムコード (`trust_remote_code`) \|
	\| `looped_gdn.py` \| モデル本体・空間マージ変換 (`python looped_gdn.py --src j-llm/Qwen3.5-2B-SpeedX2 --groups 6`) \|
	\| `scripts/colab_looped_distill.py` \| 蒸留スクリプト (Colab に貼るだけ) \|
	\| `scripts/infer_looped_gdn.py` \| 推論 / 対話スクリプト \|
	\| `scripts/colab_ast_transplant.py` \| 高度な空間移植 (活性考慮 RegMean + DP 可変長分割 + 活性考慮 LoRA) \|
	\| `scripts/test_looped_gdn.py` \| 単体テスト \|
	\| `looped_config.json` \| 旧形式の設定 (`LoopedGDNForCausalLM.load` 用) \|

	## 学習
	- Stage 1: グループ単位 hidden 蒸留 (相対 MSE + コサイン) / Stage 2: 教師 logits KL + hidden 損失
	- データ: C4 (日本語・英語)

	## 注意
	蒸留による近似で教師より性能は低下します。ライセンスは元モデル (Qwen/Qwen3.5-2B) に従います。
	元モデル: Qwen Team / 全 GDN 化: summerMC, j-llm